7.2.1 Introduction
Feature extraction is a fundamental process in biometric recognition and is a critical step towards the understanding and analysis of image content. These features are the main elements used in the matching phase. The success of a matching process depends on the accuracy of the extracted features.
This section proposes a new method for FKP recognition, which is a modification of the CLBP method. The dTCLBP feature is the combination of two components, sign and magnitude, where the sign component is the same as the original LBP while the magnitude feature is encoded by the dynamic threshold. The performance of the proposed approach has been compared to some state-of-the-art approaches.
7.2.2 A brief review of the completed local binary pattern CLBP
With its numerous properties, LBP has become a powerful discriminator. Among these properties are the fact that its computational complexity is low and it is less sensitive to changes in illumination, unlike many other descriptors [77]. Guo et al [34] used a CLBP descriptor by splitting local image differences into two complementary components, sign sp
and magnitude mp. Given a centre pixel gp and its neighbour gc, p = 0, 1, . . . , 7, the local
difference between gp and gc is defined as in Equation 7.1 [34, 76]:
sp= gp− gc, mp= |gp− gc| (7.1)
Two operators, known as CLBP S and CLBP M respectively, are proposed for coding, where the CLBP S is equivalent to the conventional LBP and the CLBP M measures the local variance of magnitude. The CLBP M can be expressed as follows [34, 76]:
CLBP MP,R= P −1 X p=0 t (mp− c)2 p (7.2) t(x, c) = 1 if x ≥ c 0 if x < c (7.3)
where the threshold c is set as the mean value of mp for the whole image. Guo et al
[34] observed that the centre pixel, which expresses the local grey level of the image, also contains discriminative information. Thus, they defined an operator known as the CLBP C to extract the local central information as follows [34, 76]:
CLBP CP,R= t (gp, cl) (7.4)
where the threshold clis set as the average grey level of the whole image. By combining the
three operators, CLBP S, CLBP M, and CLBP C, it was shown that there is a significant improvement in relation to rotation-invariant texture classification.
7.2.3 Proposed methodology: FKP recognition
A schematic diagram of the proposed FKP recognition system is shown in Figure 7.1. As mentioned previously, the system comprises the following steps: FKP image acquisition, feature extraction, reduction of features and classification. The FKP features are extracted using the dTCLBP descriptor in the first phase of recognition. The PCA is presented as a dimensionality reduction algorithm to reduce the amount of unnecessary information. Clas- sification is performed using a LDA classifier for FKP recognition. The major components of the proposed FKP system are explained below.
Figure 7.1: The general scheme of the FKP recognition system. .
7.2.3.1 dTCLBP-based feature extraction
dTCLBP is proposed based on the customized LBP texture feature for the classification of FKP images. The dTCLBP descriptor has two components which are used to encode the local texture property of an image. The first signifies the sign feature, whereas the second is the magnitude feature vector which corresponds to gradient information achieved according to the magnitude of local differences [21]. If gcis the centre pixel and gpis one of neighbours
two components, specifically sign sp and magnitude mp as defined in Equation 7.5. The local difference sign-magnitude transform is used to compute dTCLBP S and dTCLBP M, which represent texture features. dTCLBP S is the same as the original LBP descriptor. In order to encode magnitude features the threshold value λ is set to the median of the dynamic thresholds TP instead of comparing the neighbours with the mean of the absolute
value of the differences between the centre pixel and one neighbour as used in CLBP [34]. The dynamic thresholds [T0, T1, . . . , T7] are calculated as defined in Equation 7.6 to increase
the robustness of the magnitude features. As dTCLBP M is not a binary code as defined in Equation 7.7, local pattern operator is created by comparing the magnitude matrices with the median of the set of the dynamic threshold as defined in (7.8); all magnitude values greater than or equal to median threshold λ are quantized to 1, otherwise to 0 as follows:
sp = dp, mp= |dp| Radius(R) (7.5) TP= 1 − min(mp) max(mp) (7.6)
Radius R defines the distance of the neighbours to the centre, while p = 0, 1, . . . , 7 provides the number of samples for that distance (these are employed as neighbours). In addition, min(mp) is the minimum value of magnitude mp from the whole image, while max(mp)
relates to the maximum value of magnitude of mpfrom the whole image, in order to calculate
the dynamic thresholds TP for each neighbour as follows:
dTCLBP M(P,R)= P −1 X p=0 t (mp, λ) 2p (7.7) t(x, λ) = 1 if x ≥ λ 0 if x < λ (7.8)
Finally, the two different histograms, for dTCLBP S, and dTCLBP M, are concatenated to form the feature vector for pattern classification.
7.2.3.2 Feature reduction
To reduce the size of the dimension of the feature vectors of the training dataset, a multivari- ate statistical technique known as PCA is used. The basic concept of PCA was described and defined in section 2.4.1. Here, the 6 training images linked to each subject modality were employed to create the training FKP vector. As mentioned before, the training images each have a vector of FKP modality in a D × N matrix format, where D is the dTCLBP of the feature dimension and N is the total number for each modality. The covariance matrix is created using FKP training images. The principal components are obtained from an analysis that consists of finding the eigenvectors from a covariance matrix derived from a set of training images. A adequate dimensions were retained in the training FKP data to estimate 99% of the energy. Using this new d-dimensional eigenvector matrix, the training and testing FKP datasets are projected onto the new subspace.
The proposed system, the PCA stage is only used to generate the low-dimensional feature vector of the training and testing datasets separately for each type of finger modality.
7.2.3.3 Recognition stage
After reducing the size of the FKP features, the next stage is to match the features of the FKP in the testing phase with the FKP features in the training set which have the same identity. The LDA is designed to classify an FKP (unknown class) into one of the training sets (known class). The LDA approach has been employed in the considerably wider area of object recognition (such as the face, iris, or speech.). A linear border between the different classes is established for separate classes using a linear classifier. The approach adopted by Fisher [26] locates a linear combination in connection with the variables so as to separate two classes as much as possible. The most suitable possible direction is sought regarding the separation of the two classes. As discussed earlier in section 2.5.1, LDA can be achieved by maximizing the ratio of the determinant of the within-class variance and the determinant of the between-class variance. If the kth class has a Gaussian distribution with mean µk
and covariance, then the maximum conditional probability is chosen, which is equivalent to classifying as follows:
arg max
k
(gk)
where gk is the discriminant function. Just as a reminder, the definition of the discrim-
inant, as evaluated in equation 2.9, is as follows. gk(xi) = xTΣ−1k µk−
1 2µkΣ
−1
k µk+ logπk
More details on LDA classification for FKP recognition is provided in section 2.5.1.
7.2.4 Experimental results and discussion
Experiments were undertaken using the PolyU FKP database, which is publicly available. To test the performance of the proposed methodology, comparisons were made with other existing methods described in the literature (Intensity+Gabor [91], OCLPP [45] and GCRC [66]). As mentioned in section 6.3.3.1, the intensity+Gabor approach combines the ex- tracted features of intensity and Gabor the images of each FKP in a long vector [91]. Orthogonal complex locality preserving projections (OCLPP) consider distance and angle information between feature vectors to evaluate data similarity. The group collaborative representation based classifier GCRC [66] applies group-level constraints in the CRC at the classification stage.
For the results to be comparable to the reported methods, experiments were imple- mented adopting the same protocol set up, training all FKP images in the first session and testing all images in the second session (6 training sets and 6 test sets). Therefore, for each modality (LIF, LMF, RIF and RMF) there were 990 (165x6) training samples and 990 (165x6) test samples.
The results are summarized in Table 7.1. The dTCLBP offers an attractive recognition performance rate of 92.12% in relation to LIF. Moreover, it can be seen that in the case of LMF, the dTCLBP method achieves the best result with a rate of 93.03% as shown in Figures 7.2 and 7.3. Furthermore, if compared against other existing methods, it results in increased performance of 4.44%, 5.54% and 2.63% in contrast to ’Intensity+Gabor’, OCLPP and GCRC respectively. With regards to RIF, the results displayed in the table
Table 7.1: Recognition rate precentages for the proposed method with similar state-of- the-art methods using PolyU FKP database with 6 samples of training and 6 samples of testing.
Method Recognition Rate (%)
LIF LMF RIF RMF
Intensity+Gabor [91] 89.90 88.59 89.49 88.48
OCLPP [45] 87.87 87.49 86.94 87.38
GCRC [66] 90.51 90.40 91.01 91.01
Proposed method (dTCLBP) 92.12 93.03 91.51 92.02
Figure 7.2: Comparison of proposed dTCLBP approach where 6 images for each type of FKPs (LIF, LMF, RIF, and RMF) were selected for training and the rest of the images used as test images
.
demonstrate that the dTCLBP method achieves an accuracy of 91.51%, which is higher than the ’Intensity+Gabor’, OCLPP and GCRC methods. In the case of of RMF, dTCLBP achieves the best result with a performance accuracy of 92.02%, which is also higher than the existing ’Intensity+Gabor’, OCLPP and GCRC methods.
Figure 7.3: Comparison of proposed dTCLBP method for different modalities (LIF, LMF, RIF and RMF)
.